How Generative AI Impacts Writing and Storytelling
Spoiler: Not positively.
TECHWRITING
There is so much that can be said on the topic of Large Language Models (LLMs) like Claude, ChatGPT, DeepSeek, and Gemini. Endless and important discussions can be had on the financial implications, negative environmental impact, declining critical thinking ability in users, theft of intellectual property and artistic skill, casual dismissal of privacy, and on and on and on I can go. However, as a book editor and a supporter of creatively written narratives, generative AI is problematic for me in a very specific way.
First, it is important to understand what generative language systems actually do. In the simplest of terms, data sets—all of the text available on the internet—are fed to processors, the processors learn the patterns and repeated relationships between the points of data, and then, when prompted by a request, the processors are able to generate a "new" bit of data using the data sets they were trained on. What is generated is based on probability and prediction of the kinds of words, phrases, and sequences that most often follow each other in the context of the prompt.
What this means is, if you type the prompt "write me a story about a young girl and a dragonfly" into an LLM, it is very unlikely that the LLM will come up with a dark story where the dragonfly ends up enslaving the girl. It is highly likely that the story will be aimed at younger children and that there will be some kind of easy moral at the end—because this is the pattern most stories about young girls follow. So, LLMs are not innovating, they are repackaging and regurgitating. They are "stochastic parrots," probability programs that pump out patterns of words without actually comprehending their meaning.
Now to the heart of the matter—it is no secret that organizations that are developing their own "AIs" are stealing the data they use to train these models. For instance, I used generative AI to create the feature image for this article. The robot hand writing in a journal was generated based on real artists' creations. The hand was clearly lifted from a different graphic than the flowers, and the journal and sun-reflecting river are also from different pieces of art. The art generating processor that mashed this image together uses the same predictive principles as the fancy chatbots do—the image processing algorithm is predicting which pixels most likely fit together based on other images of a similar style.
For its purpose, the image above is passable. However, a human artist would have created a graphic that is more uniform, more poignant perhaps, or less on the nose, and certainly a tad more original. It is the same when it comes to writing. Whether it is fiction or non-fiction, narratives have memorable impact on the reader because of the author's approach to the story—how they feel about the topic, how their characters choose to deal with traumatic events, how their curiosity and passion informs the information they deliver. If a writer chooses to use generative AI to create part of a draft, do their research, or even create an outline for the piece, that's great, they got the project started! But I will always (gently) question, "How much of you is in there?"
Why does this matter? It matters because language models that can only create approximations of philosophical arguments or research-based critiques of government or fantasy worlds and magic systems based on probability algorithms will, invariably, create flatter, featureless, and forgettable narratives. Prediction algorithms use existing data to generate whatever you ask them to and existing data is flawed. Existing data is biased. Existing data skews toward whatever there is the most of. As a result, when prompting "write me a story about a young girl and a dragonfly," the young girl in the story will most certainly not be a person of color, will most certainly have heterosexual parents, will most certainly be good-natured, and the dragonfly will most certainly cross paths with the girl on a warm and sunny day.
Nuance through specificity is what captures the attention of a diverse human mind. Due to how they function, LLM's are unable to approximate these opposing ideas effectively and are therefore unable to "write" impactful narratives. Narratives have true impact on readers when they are rooted in the realities of emotion and the human experience. The invariable variance in perception, feelings, attitude, experience, and opinion is what a writer brings to their work and these are the ingredients that make the resulting writing layered enough for a reader to enjoy. Handing over the job to a predictive language model without intervention results in the proliferation of flaws and inherent biases and the dwindling of truly human-centered and reader-centered narratives.
The process by which we create and tell stories will continue to change due to technological advances, as it has done repeatedly in the past. What will not change is what readers find fulfilling—connection to the vast minutiae of the human experience through the writer's words. By its nature, generative AI strips that connection away, adding a distance littered with predictions and probabilities between the writer and the reader.
As a book editor, I help writers to strengthen the impact their narratives will have on readers by building in specific aspects of the human experience. So a predictive algorithm that tends to make all the young girls in stories white and all the dragonflies shimmer in the sun as they espouse generic moral principles? Concerning indeed.
